hammer2_chain_t *
hammer2_inode_chain_and_parent(hammer2_inode_t *ip, int clindex,
hammer2_chain_t **parentp, int how)
{
hammer2_chain_t *chain;
hammer2_chain_t *parent;
for (;;) {
hammer2_spin_sh(&ip->cluster_spin);
if (clindex >= ip->cluster.nchains)
chain = NULL;
else
chain = ip->cluster.array[clindex].chain;
if (chain) {
hammer2_chain_ref(chain);
hammer2_spin_unsh(&ip->cluster_spin);
hammer2_chain_lock(chain, how);
} else {
hammer2_spin_unsh(&ip->cluster_spin);
}
/*
* Get parent, lock order must be (parent, chain).
*/
parent = chain->parent;
if (parent) {
hammer2_chain_ref(parent);
hammer2_chain_unlock(chain);
hammer2_chain_lock(parent, how);
hammer2_chain_lock(chain, how);
}
if (ip->cluster.array[clindex].chain == chain &&
chain->parent == parent) {
break;
}
/*
* Retry
*/
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
}
*parentp = parent;
return chain;
}
static int
hammer2_ioctl_pfs_snapshot(hammer2_inode_t *ip, void *data)
{
hammer2_ioc_pfs_t *pfs = data;
hammer2_dev_t *hmp;
hammer2_chain_t *chain;
hammer2_tid_t mtid;
int error;
if (pfs->name[0] == 0)
return(EINVAL);
if (pfs->name[sizeof(pfs->name)-1] != 0)
return(EINVAL);
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
hammer2_vfs_sync(ip->pmp->mp, MNT_WAIT);
hammer2_trans_init(ip->pmp, HAMMER2_TRANS_ISFLUSH);
mtid = hammer2_trans_sub(ip->pmp);
hammer2_inode_lock(ip, 0);
chain = hammer2_inode_chain(ip, 0, HAMMER2_RESOLVE_ALWAYS);
error = hammer2_chain_snapshot(chain, pfs, mtid);
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
hammer2_inode_unlock(ip);
hammer2_trans_done(ip->pmp);
return (error);
}
/*
* Repoint ip->chain to nchain. Caller must hold the inode exclusively
* locked.
*
* ip->chain is set to nchain. The prior chain in ip->chain is dropped
* and nchain is ref'd.
*/
void
hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
hammer2_chain_t *nchain)
{
hammer2_chain_t *ochain;
hammer2_inode_t *opip;
/*
* Repoint ip->chain if requested.
*/
ochain = ip->chain;
ip->chain = nchain;
if (nchain)
hammer2_chain_ref(nchain);
if (ochain)
hammer2_chain_drop(ochain);
/*
* Repoint ip->pip if requested (non-NULL pip).
*/
if (pip && ip->pip != pip) {
opip = ip->pip;
hammer2_inode_ref(pip);
ip->pip = pip;
if (opip)
hammer2_inode_drop(opip);
}
}
/*
* Locate next match or overlap under parent, replace cluster
*/
hammer2_cluster_t *
hammer2_cluster_next(hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
hammer2_key_t *key_nextp,
hammer2_key_t key_beg, hammer2_key_t key_end, int flags)
{
hammer2_chain_t *chain;
hammer2_key_t key_accum;
hammer2_key_t key_next;
int null_count;
int i;
key_accum = *key_nextp;
null_count = 0;
cluster->focus = NULL;
cparent->focus = NULL;
for (i = 0; i < cparent->nchains; ++i) {
key_next = *key_nextp;
chain = cluster->array[i];
if (chain == NULL) {
if (cparent->focus == NULL)
cparent->focus = cparent->array[i];
++null_count;
continue;
}
if (cparent->array[i] == NULL) {
if (flags & HAMMER2_LOOKUP_NOLOCK)
hammer2_chain_drop(chain);
else
hammer2_chain_unlock(chain);
++null_count;
continue;
}
chain = hammer2_chain_next(&cparent->array[i], chain,
&key_next, key_beg, key_end,
&cparent->cache_index[i], flags);
if (cparent->focus == NULL)
cparent->focus = cparent->array[i];
cluster->array[i] = chain;
if (chain == NULL) {
++null_count;
} else if (cluster->focus == NULL) {
cluster->focus = chain;
}
if (key_accum > key_next)
key_accum = key_next;
}
if (null_count == i) {
hammer2_cluster_drop(cluster);
cluster = NULL;
}
return(cluster);
}
/*
* Destroy an extranious chain.
*
* Both *parentp and *chainp are locked shared.
*
* On return, *chainp will be adjusted to point to the next element in the
* iteration and locked shared.
*/
static
int
hammer2_sync_destroy(hammer2_thread_t *thr,
hammer2_chain_t **parentp, hammer2_chain_t **chainp,
hammer2_tid_t mtid, int idx)
{
hammer2_chain_t *chain;
hammer2_chain_t *parent;
hammer2_key_t key_next;
hammer2_key_t save_key;
int cache_index = -1;
chain = *chainp;
#if HAMMER2_THREAD_DEBUG
if (hammer2_debug & 1)
kprintf("destroy rec %p/%p slave %d %d.%016jx\n",
*parentp, chain,
idx, chain->bref.type, chain->bref.key);
#endif
save_key = chain->bref.key;
if (save_key != HAMMER2_KEY_MAX)
++save_key;
/*
* Try to avoid unnecessary I/O.
*
* XXX accounting not propagated up properly. We might have to do
* a RESOLVE_MAYBE here and pass 0 for the flags.
*/
hammer2_chain_unlock(chain); /* relock exclusive */
hammer2_chain_unlock(*parentp);
hammer2_chain_lock(*parentp, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_NEVER);
hammer2_chain_delete(*parentp, chain, mtid, HAMMER2_DELETE_PERMANENT);
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
chain = NULL; /* safety */
hammer2_chain_unlock(*parentp); /* relock shared */
hammer2_chain_lock(*parentp, HAMMER2_RESOLVE_SHARED |
HAMMER2_RESOLVE_ALWAYS);
*chainp = hammer2_chain_lookup(&parent, &key_next,
save_key, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED |
HAMMER2_LOOKUP_NODIRECT |
HAMMER2_LOOKUP_NODATA);
return 0;
}
/*
* HAMMER2 inode locks
*
* HAMMER2 offers shared locks and exclusive locks on inodes.
*
* An inode's ip->chain pointer is resolved and stable while an inode is
* locked, and can be cleaned out at any time (become NULL) when an inode
* is not locked.
*
* This function handles duplication races and hardlink replacement races
* which can cause ip's cached chain to become stale.
*
* The underlying chain is also locked and returned.
*
* NOTE: We don't combine the inode/chain lock because putting away an
* inode would otherwise confuse multiple lock holders of the inode.
*/
hammer2_chain_t *
hammer2_inode_lock_ex(hammer2_inode_t *ip)
{
hammer2_chain_t *chain;
hammer2_chain_t *ochain;
hammer2_chain_core_t *core;
int error;
hammer2_inode_ref(ip);
ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
chain = ip->chain;
core = chain->core;
for (;;) {
if (chain->flags & HAMMER2_CHAIN_DUPLICATED) {
spin_lock(&core->cst.spin);
while (chain->flags & HAMMER2_CHAIN_DUPLICATED)
chain = TAILQ_NEXT(chain, core_entry);
hammer2_chain_ref(chain);
spin_unlock(&core->cst.spin);
hammer2_inode_repoint(ip, NULL, chain);
hammer2_chain_drop(chain);
}
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
if ((chain->flags & HAMMER2_CHAIN_DUPLICATED) == 0)
break;
hammer2_chain_unlock(chain);
}
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK &&
(chain->flags & HAMMER2_CHAIN_DELETED) == 0) {
error = hammer2_hardlink_find(ip->pip, &chain, &ochain);
hammer2_chain_drop(ochain);
KKASSERT(error == 0);
/* XXX error handling */
}
return (chain);
}
/*
* HAMMER2 inode locks
*
* HAMMER2 offers shared locks and exclusive locks on inodes.
*
* An inode's ip->chain pointer is resolved and stable while an inode is
* locked, and can be cleaned out at any time (become NULL) when an inode
* is not locked.
*
* The underlying chain is also locked and returned.
*
* NOTE: We don't combine the inode/chain lock because putting away an
* inode would otherwise confuse multiple lock holders of the inode.
*/
hammer2_chain_t *
hammer2_inode_lock_ex(hammer2_inode_t *ip)
{
hammer2_chain_t *chain;
hammer2_inode_ref(ip);
ccms_thread_lock(&ip->topo_cst, CCMS_STATE_EXCLUSIVE);
/*
* ip->chain fixup. Certain duplications used to move inodes
* into indirect blocks (for example) can cause ip->chain to
* become stale.
*/
again:
chain = ip->chain;
if (hammer2_chain_refactor_test(chain, 1)) {
spin_lock(&chain->core->cst.spin);
while (hammer2_chain_refactor_test(chain, 1))
chain = chain->next_parent;
if (ip->chain != chain) {
hammer2_chain_ref(chain);
spin_unlock(&chain->core->cst.spin);
hammer2_inode_repoint(ip, NULL, chain);
hammer2_chain_drop(chain);
} else {
spin_unlock(&chain->core->cst.spin);
}
}
KKASSERT(chain != NULL); /* for now */
hammer2_chain_lock(chain, HAMMER2_RESOLVE_ALWAYS);
/*
* Resolve duplication races
*/
if (hammer2_chain_refactor_test(chain, 1)) {
hammer2_chain_unlock(chain);
goto again;
}
return (chain);
}
/*
* Drop the caller's reference to the cluster. When the ref count drops to
* zero this function frees the cluster and drops all underlying chains.
*/
void
hammer2_cluster_drop(hammer2_cluster_t *cluster)
{
hammer2_chain_t *chain;
int i;
KKASSERT(cluster->refs > 0);
for (i = 0; i < cluster->nchains; ++i) {
chain = cluster->array[i];
if (chain) {
hammer2_chain_drop(chain);
if (cluster->refs == 1)
cluster->array[i] = NULL;
}
}
if (atomic_fetchadd_int(&cluster->refs, -1) == 1) {
cluster->focus = NULL;
kfree(cluster, M_HAMMER2);
/* cluster = NULL; safety */
}
}
/*
* Repoint a single element from the cluster to the ip. Used by the
* synchronization threads to piecemeal update inodes. Does not change
* focus and requires inode to be re-locked to clean-up flags (XXX).
*/
void
hammer2_inode_repoint_one(hammer2_inode_t *ip, hammer2_cluster_t *cluster,
int idx)
{
hammer2_chain_t *ochain;
hammer2_chain_t *nchain;
int i;
hammer2_spin_ex(&ip->cluster_spin);
KKASSERT(idx < cluster->nchains);
if (idx < ip->cluster.nchains) {
ochain = ip->cluster.array[idx].chain;
nchain = cluster->array[idx].chain;
} else {
ochain = NULL;
nchain = cluster->array[idx].chain;
for (i = ip->cluster.nchains; i <= idx; ++i) {
bzero(&ip->cluster.array[i],
sizeof(ip->cluster.array[i]));
ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID;
}
ip->cluster.nchains = idx + 1;
}
if (ochain != nchain) {
/*
* Make adjustments.
*/
ip->cluster.array[idx].chain = nchain;
ip->cluster.array[idx].flags &= ~HAMMER2_CITEM_INVALID;
ip->cluster.array[idx].flags |= cluster->array[idx].flags &
HAMMER2_CITEM_INVALID;
}
hammer2_spin_unex(&ip->cluster_spin);
if (ochain != nchain) {
if (nchain)
hammer2_chain_ref(nchain);
if (ochain)
hammer2_chain_drop(ochain);
}
}
/*
* XXX initial NULL cluster needs reworking (pass **clusterp ?)
*
* The raw scan function is similar to lookup/next but does not seek to a key.
* Blockrefs are iterated via first_chain = (parent, NULL) and
* next_chain = (parent, chain).
*
* The passed-in parent must be locked and its data resolved. The returned
* chain will be locked. Pass chain == NULL to acquire the first sub-chain
* under parent and then iterate with the passed-in chain (which this
* function will unlock).
*/
hammer2_cluster_t *
hammer2_cluster_scan(hammer2_cluster_t *cparent, hammer2_cluster_t *cluster,
int flags)
{
hammer2_chain_t *chain;
int null_count;
int i;
null_count = 0;
for (i = 0; i < cparent->nchains; ++i) {
chain = cluster->array[i];
if (chain == NULL) {
++null_count;
continue;
}
if (cparent->array[i] == NULL) {
if (flags & HAMMER2_LOOKUP_NOLOCK)
hammer2_chain_drop(chain);
else
hammer2_chain_unlock(chain);
++null_count;
continue;
}
chain = hammer2_chain_scan(cparent->array[i], chain,
&cparent->cache_index[i], flags);
cluster->array[i] = chain;
if (chain == NULL)
++null_count;
}
if (null_count == i) {
hammer2_cluster_drop(cluster);
cluster = NULL;
}
return(cluster);
}
/*
* Unlink the file from the specified directory inode. The directory inode
* does not need to be locked.
*
* isdir determines whether a directory/non-directory check should be made.
* No check is made if isdir is set to -1.
*/
int
hammer2_unlink_file(hammer2_inode_t *dip,
const uint8_t *name, size_t name_len,
int isdir, hammer2_inode_t *retain_ip)
{
hammer2_mount_t *hmp;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_chain_t *dparent;
hammer2_chain_t *dchain;
hammer2_key_t lhc;
hammer2_inode_t *ip;
hammer2_inode_t *oip;
int error;
uint8_t type;
error = 0;
oip = NULL;
hmp = dip->hmp;
lhc = hammer2_dirhash(name, name_len);
/*
* Search for the filename in the directory
*/
parent = &dip->chain;
hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
chain = hammer2_chain_lookup(hmp, &parent,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
0);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
chain->u.ip &&
name_len == chain->data->ipdata.name_len &&
bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
break;
}
chain = hammer2_chain_next(hmp, &parent, chain,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
0);
}
/*
* Not found or wrong type (isdir < 0 disables the type check).
*/
if (chain == NULL) {
hammer2_chain_unlock(hmp, parent);
return ENOENT;
}
if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK)
type = chain->data->ipdata.target_type;
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
error = ENOTDIR;
goto done;
}
if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
error = EISDIR;
goto done;
}
/*
* Hardlink must be resolved. We can't hold parent locked while we
* do this or we could deadlock.
*/
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
hammer2_chain_unlock(hmp, parent);
parent = NULL;
error = hammer2_hardlink_find(dip, &chain, &oip);
}
/*
* If this is a directory the directory must be empty. However, if
* isdir < 0 we are doing a rename and the directory does not have
* to be empty.
*
* NOTE: We check the full key range here which covers both visible
* and invisible entries. Theoretically there should be no
* invisible (hardlink target) entries if there are no visible
* entries.
*/
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir >= 0) {
dparent = chain;
hammer2_chain_lock(hmp, dparent, HAMMER2_RESOLVE_ALWAYS);
dchain = hammer2_chain_lookup(hmp, &dparent,
0, (hammer2_key_t)-1,
HAMMER2_LOOKUP_NODATA);
if (dchain) {
hammer2_chain_unlock(hmp, dchain);
hammer2_chain_unlock(hmp, dparent);
error = ENOTEMPTY;
goto done;
}
hammer2_chain_unlock(hmp, dparent);
dparent = NULL;
/* dchain NULL */
}
/*
* Ok, we can now unlink the chain. We always decrement nlinks even
* if the entry can be deleted in case someone has the file open and
* does an fstat().
*
* The chain itself will no longer be in the on-media topology but
* can still be flushed to the media (e.g. if an open descriptor
* remains). When the last vnode/ip ref goes away the chain will
* be marked unmodified, avoiding any further (now unnecesary) I/O.
*/
if (oip) {
/*
* If this was a hardlink we first delete the hardlink
* pointer entry.
*/
parent = oip->chain.parent;
hammer2_chain_lock(hmp, parent, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(hmp, &oip->chain, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_delete(hmp, parent, &oip->chain,
(retain_ip == oip));
hammer2_chain_unlock(hmp, &oip->chain);
hammer2_chain_unlock(hmp, parent);
parent = NULL;
/*
* Then decrement nlinks on hardlink target.
*/
ip = chain->u.ip;
if (ip->ip_data.nlinks == 1) {
dparent = chain->parent;
hammer2_chain_ref(hmp, chain);
hammer2_chain_unlock(hmp, chain);
hammer2_chain_lock(hmp, dparent,
HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(hmp, chain, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_drop(hmp, chain);
hammer2_chain_modify(hmp, chain, 0);
--ip->ip_data.nlinks;
hammer2_chain_delete(hmp, dparent, chain, 0);
hammer2_chain_unlock(hmp, dparent);
} else {
hammer2_chain_modify(hmp, chain, 0);
--ip->ip_data.nlinks;
}
} else {
/*
* Otherwise this was not a hardlink and we can just
* remove the entry and decrement nlinks.
*/
ip = chain->u.ip;
hammer2_chain_modify(hmp, chain, 0);
--ip->ip_data.nlinks;
hammer2_chain_delete(hmp, parent, chain,
(retain_ip == ip));
}
error = 0;
done:
if (chain)
hammer2_chain_unlock(hmp, chain);
if (parent)
hammer2_chain_unlock(hmp, parent);
if (oip)
hammer2_chain_drop(oip->hmp, &oip->chain);
return error;
}
/*
* (Frontend) collect a response from a running cluster op.
*
* Responses are fed from all appropriate nodes concurrently
* and collected into a cohesive response >= collect_key.
*
* The collector will return the instant quorum or other requirements
* are met, even if some nodes get behind or become non-responsive.
*
* HAMMER2_XOP_COLLECT_NOWAIT - Used to 'poll' a completed collection,
* usually called synchronously from the
* node XOPs for the strategy code to
* fake the frontend collection and complete
* the BIO as soon as possible.
*
* HAMMER2_XOP_SYNCHRONIZER - Reqeuest synchronization with a particular
* cluster index, prevents looping when that
* index is out of sync so caller can act on
* the out of sync element. ESRCH and EDEADLK
* can be returned if this flag is specified.
*
* Returns 0 on success plus a filled out xop->cluster structure.
* Return ENOENT on normal termination.
* Otherwise return an error.
*/
int
hammer2_xop_collect(hammer2_xop_head_t *xop, int flags)
{
hammer2_xop_fifo_t *fifo;
hammer2_chain_t *chain;
hammer2_key_t lokey;
int error;
int keynull;
int adv; /* advance the element */
int i;
uint32_t check_counter;
loop:
/*
* First loop tries to advance pieces of the cluster which
* are out of sync.
*/
lokey = HAMMER2_KEY_MAX;
keynull = HAMMER2_CHECK_NULL;
check_counter = xop->check_counter;
cpu_lfence();
for (i = 0; i < xop->cluster.nchains; ++i) {
chain = xop->cluster.array[i].chain;
if (chain == NULL) {
adv = 1;
} else if (chain->bref.key < xop->collect_key) {
adv = 1;
} else {
keynull &= ~HAMMER2_CHECK_NULL;
if (lokey > chain->bref.key)
lokey = chain->bref.key;
adv = 0;
}
if (adv == 0)
continue;
/*
* Advance element if possible, advanced element may be NULL.
*/
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
fifo = &xop->collect[i];
if (fifo->ri != fifo->wi) {
cpu_lfence();
chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK];
++fifo->ri;
xop->cluster.array[i].chain = chain;
if (chain == NULL) {
/* XXX */
xop->cluster.array[i].flags |=
HAMMER2_CITEM_NULL;
}
if (fifo->wi - fifo->ri < HAMMER2_XOPFIFO / 2)
wakeup(xop); /* XXX optimize */
--i; /* loop on same index */
} else {
/*
* Retain CITEM_NULL flag. If set just repeat EOF.
* If not, the NULL,0 combination indicates an
* operation in-progress.
*/
xop->cluster.array[i].chain = NULL;
/* retain any CITEM_NULL setting */
}
}
/*
* Determine whether the lowest collected key meets clustering
* requirements. Returns:
*
* 0 - key valid, cluster can be returned.
*
* ENOENT - normal end of scan, return ENOENT.
*
* ESRCH - sufficient elements collected, quorum agreement
* that lokey is not a valid element and should be
* skipped.
*
* EDEADLK - sufficient elements collected, no quorum agreement
* (and no agreement possible). In this situation a
* repair is needed, for now we loop.
*
* EINPROGRESS - insufficient elements collected to resolve, wait
* for event and loop.
*/
if ((flags & HAMMER2_XOP_COLLECT_WAITALL) &&
xop->run_mask != HAMMER2_XOPMASK_VOP) {
error = EINPROGRESS;
} else {
error = hammer2_cluster_check(&xop->cluster, lokey, keynull);
}
if (error == EINPROGRESS) {
if (xop->check_counter == check_counter) {
if (flags & HAMMER2_XOP_COLLECT_NOWAIT)
goto done;
tsleep_interlock(&xop->check_counter, 0);
cpu_lfence();
if (xop->check_counter == check_counter) {
tsleep(&xop->check_counter, PINTERLOCKED,
"h2coll", hz*60);
}
}
goto loop;
}
if (error == ESRCH) {
if (lokey != HAMMER2_KEY_MAX) {
xop->collect_key = lokey + 1;
goto loop;
}
error = ENOENT;
}
if (error == EDEADLK) {
kprintf("hammer2: no quorum possible lokey %016jx\n",
lokey);
if (lokey != HAMMER2_KEY_MAX) {
xop->collect_key = lokey + 1;
goto loop;
}
error = ENOENT;
}
if (lokey == HAMMER2_KEY_MAX)
xop->collect_key = lokey;
else
xop->collect_key = lokey + 1;
done:
return error;
}
/*
* Find a specific PFS by name
*/
static int
hammer2_ioctl_pfs_lookup(hammer2_inode_t *ip, void *data)
{
const hammer2_inode_data_t *ripdata;
hammer2_dev_t *hmp;
hammer2_ioc_pfs_t *pfs;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_key_t key_next;
hammer2_key_t lhc;
int cache_index = -1;
int error;
size_t len;
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
pfs = data;
error = 0;
hammer2_inode_lock(hmp->spmp->iroot, HAMMER2_RESOLVE_SHARED);
parent = hammer2_inode_chain(hmp->spmp->iroot, 0,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
pfs->name[sizeof(pfs->name) - 1] = 0;
len = strlen(pfs->name);
lhc = hammer2_dirhash(pfs->name, len);
chain = hammer2_chain_lookup(&parent, &key_next,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index,
HAMMER2_LOOKUP_SHARED);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE) {
ripdata = &chain->data->ipdata;
if (ripdata->meta.name_len == len &&
bcmp(ripdata->filename, pfs->name, len) == 0) {
break;
}
ripdata = NULL; /* safety */
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next,
lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index,
HAMMER2_LOOKUP_SHARED);
}
/*
* Load the data being returned by the ioctl.
*/
if (chain) {
ripdata = &chain->data->ipdata;
pfs->name_key = ripdata->meta.name_key;
pfs->pfs_type = ripdata->meta.pfs_type;
pfs->pfs_subtype = ripdata->meta.pfs_subtype;
pfs->pfs_clid = ripdata->meta.pfs_clid;
pfs->pfs_fsid = ripdata->meta.pfs_fsid;
ripdata = NULL;
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
} else {
error = ENOENT;
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
hammer2_inode_unlock(hmp->spmp->iroot);
return (error);
}
/*
* Create a new PFS under the super-root
*/
static int
hammer2_ioctl_pfs_create(hammer2_inode_t *ip, void *data)
{
hammer2_inode_data_t *nipdata;
hammer2_chain_t *nchain;
hammer2_dev_t *hmp;
hammer2_ioc_pfs_t *pfs;
hammer2_inode_t *nip;
hammer2_tid_t mtid;
int error;
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
pfs = data;
nip = NULL;
if (pfs->name[0] == 0)
return(EINVAL);
pfs->name[sizeof(pfs->name) - 1] = 0; /* ensure 0-termination */
if (hammer2_ioctl_pfs_lookup(ip, pfs) == 0)
return(EEXIST);
hammer2_trans_init(hmp->spmp, 0);
mtid = hammer2_trans_sub(hmp->spmp);
nip = hammer2_inode_create(hmp->spmp->iroot, NULL, NULL,
pfs->name, strlen(pfs->name), 0,
1, HAMMER2_OBJTYPE_DIRECTORY, 0,
HAMMER2_INSERT_PFSROOT, &error);
if (error == 0) {
hammer2_inode_modify(nip);
nchain = hammer2_inode_chain(nip, 0, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_modify(nchain, mtid, 0);
nipdata = &nchain->data->ipdata;
nip->meta.pfs_type = pfs->pfs_type;
nip->meta.pfs_subtype = pfs->pfs_subtype;
nip->meta.pfs_clid = pfs->pfs_clid;
nip->meta.pfs_fsid = pfs->pfs_fsid;
nip->meta.op_flags |= HAMMER2_OPFLAG_PFSROOT;
/*
* Set default compression and check algorithm. This
* can be changed later.
*
* Do not allow compression on PFS's with the special name
* "boot", the boot loader can't decompress (yet).
*/
nip->meta.comp_algo =
HAMMER2_ENC_ALGO(HAMMER2_COMP_NEWFS_DEFAULT);
nip->meta.check_algo =
HAMMER2_ENC_ALGO( HAMMER2_CHECK_ISCSI32);
if (strcasecmp(pfs->name, "boot") == 0) {
nip->meta.comp_algo =
HAMMER2_ENC_ALGO(HAMMER2_COMP_AUTOZERO);
}
#if 0
hammer2_blockref_t bref;
/* XXX new PFS needs to be rescanned / added */
bref = nchain->bref;
kprintf("ADD LOCAL PFS (IOCTL): %s\n", nipdata->filename);
hammer2_pfsalloc(nchain, nipdata, bref.modify_tid);
#endif
/* XXX rescan */
hammer2_chain_unlock(nchain);
hammer2_chain_drop(nchain);
/*
* Super-root isn't mounted, fsync it
*/
hammer2_inode_ref(nip);
hammer2_inode_unlock(nip);
hammer2_inode_fsync(nip);
hammer2_inode_drop(nip);
}
hammer2_trans_done(hmp->spmp);
return (error);
}
void
hammer2_inode_unlock_nlinks(hammer2_inode_t *ip)
{
hammer2_chain_drop(ip->hmp, &ip->chain);
}
/*
* Used to scan and retrieve PFS information. PFS's are directories under
* the super-root.
*
* To scan PFSs pass name_key=0. The function will scan for the next
* PFS and set all fields, as well as set name_next to the next key.
* When no PFSs remain, name_next is set to (hammer2_key_t)-1.
*
* To retrieve a particular PFS by key, specify the key but note that
* the ioctl will return the lowest key >= specified_key, so the caller
* must verify the key.
*
* To retrieve the PFS associated with the file descriptor, pass
* name_key set to (hammer2_key_t)-1.
*/
static int
hammer2_ioctl_pfs_get(hammer2_inode_t *ip, void *data)
{
const hammer2_inode_data_t *ripdata;
hammer2_dev_t *hmp;
hammer2_ioc_pfs_t *pfs;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_key_t key_next;
hammer2_key_t save_key;
int cache_index = -1;
int error;
hmp = ip->pmp->pfs_hmps[0];
if (hmp == NULL)
return (EINVAL);
pfs = data;
save_key = pfs->name_key;
error = 0;
/*
* Setup
*/
if (save_key == (hammer2_key_t)-1) {
hammer2_inode_lock(ip->pmp->iroot, 0);
parent = NULL;
chain = hammer2_inode_chain(hmp->spmp->iroot, 0,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
} else {
hammer2_inode_lock(hmp->spmp->iroot, 0);
parent = hammer2_inode_chain(hmp->spmp->iroot, 0,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
chain = hammer2_chain_lookup(&parent, &key_next,
pfs->name_key, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED);
}
/*
* Locate next PFS
*/
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE)
break;
if (parent == NULL) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
chain = NULL;
break;
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED);
}
/*
* Load the data being returned by the ioctl.
*/
if (chain) {
ripdata = &chain->data->ipdata;
pfs->name_key = ripdata->meta.name_key;
pfs->pfs_type = ripdata->meta.pfs_type;
pfs->pfs_subtype = ripdata->meta.pfs_subtype;
pfs->pfs_clid = ripdata->meta.pfs_clid;
pfs->pfs_fsid = ripdata->meta.pfs_fsid;
KKASSERT(ripdata->meta.name_len < sizeof(pfs->name));
bcopy(ripdata->filename, pfs->name, ripdata->meta.name_len);
pfs->name[ripdata->meta.name_len] = 0;
ripdata = NULL; /* safety */
/*
* Calculate name_next, if any.
*/
if (parent == NULL) {
pfs->name_next = (hammer2_key_t)-1;
} else {
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED);
if (chain)
pfs->name_next = chain->bref.key;
else
pfs->name_next = (hammer2_key_t)-1;
}
} else {
pfs->name_next = (hammer2_key_t)-1;
error = ENOENT;
}
/*
* Cleanup
*/
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
if (save_key == (hammer2_key_t)-1) {
hammer2_inode_unlock(ip->pmp->iroot);
} else {
hammer2_inode_unlock(hmp->spmp->iroot);
}
return (error);
}
/*
* Each out of sync node sync-thread must issue an all-nodes XOP scan of
* the inode. This creates a multiplication effect since the XOP scan itself
* issues to all nodes. However, this is the only way we can safely
* synchronize nodes which might have disparate I/O bandwidths and the only
* way we can safely deal with stalled nodes.
*/
static
int
hammer2_sync_slaves(hammer2_thread_t *thr, hammer2_inode_t *ip,
hammer2_deferred_list_t *list)
{
hammer2_xop_scanall_t *xop;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_pfs_t *pmp;
hammer2_key_t key_next;
hammer2_tid_t sync_tid;
int cache_index = -1;
int needrescan;
int wantupdate;
int error;
int nerror;
int idx;
int n;
pmp = ip->pmp;
idx = thr->clindex; /* cluster node we are responsible for */
needrescan = 0;
wantupdate = 0;
if (ip->cluster.focus == NULL)
return (EINPROGRESS);
sync_tid = ip->cluster.focus->bref.modify_tid;
#if 0
/*
* Nothing to do if all slaves are synchronized.
* Nothing to do if cluster not authoritatively readable.
*/
if (pmp->cluster_flags & HAMMER2_CLUSTER_SSYNCED)
return(0);
if ((pmp->cluster_flags & HAMMER2_CLUSTER_RDHARD) == 0)
return(HAMMER2_ERROR_INCOMPLETE);
#endif
error = 0;
/*
* The inode is left unlocked during the scan. Issue a XOP
* that does *not* include our cluster index to iterate
* properly synchronized elements and resolve our cluster index
* against it.
*/
hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
xop = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
xop->key_beg = HAMMER2_KEY_MIN;
xop->key_end = HAMMER2_KEY_MAX;
hammer2_xop_start_except(&xop->head, hammer2_xop_scanall, idx);
parent = hammer2_inode_chain(ip, idx,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
if (parent->bref.modify_tid != sync_tid)
wantupdate = 1;
hammer2_inode_unlock(ip);
chain = hammer2_chain_lookup(&parent, &key_next,
HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED |
HAMMER2_LOOKUP_NODIRECT |
HAMMER2_LOOKUP_NODATA);
error = hammer2_xop_collect(&xop->head, 0);
kprintf("XOP_INITIAL xop=%p clindex %d on %s\n", xop, thr->clindex,
pmp->pfs_names[thr->clindex]);
for (;;) {
/*
* We are done if our scan is done and the XOP scan is done.
* We are done if the XOP scan failed (that is, we don't
* have authoritative data to synchronize with).
*/
int advance_local = 0;
int advance_xop = 0;
int dodefer = 0;
hammer2_chain_t *focus;
kprintf("loop xop=%p chain[1]=%p lockcnt=%d\n",
xop, xop->head.cluster.array[1].chain,
(xop->head.cluster.array[1].chain ?
xop->head.cluster.array[1].chain->lockcnt : -1)
);
if (chain == NULL && error == ENOENT)
break;
if (error && error != ENOENT)
break;
/*
* Compare
*/
if (chain && error == ENOENT) {
/*
* If we have local chains but the XOP scan is done,
* the chains need to be deleted.
*/
n = -1;
focus = NULL;
} else if (chain == NULL) {
/*
* If our local scan is done but the XOP scan is not,
* we need to create the missing chain(s).
*/
n = 1;
focus = xop->head.cluster.focus;
} else {
/*
* Otherwise compare to determine the action
* needed.
*/
focus = xop->head.cluster.focus;
n = hammer2_chain_cmp(chain, focus);
}
/*
* Take action based on comparison results.
*/
if (n < 0) {
/*
* Delete extranious local data. This will
* automatically advance the chain.
*/
nerror = hammer2_sync_destroy(thr, &parent, &chain,
0, idx);
} else if (n == 0 && chain->bref.modify_tid !=
focus->bref.modify_tid) {
/*
* Matching key but local data or meta-data requires
* updating. If we will recurse, we still need to
* update to compatible content first but we do not
* synchronize modify_tid until the entire recursion
* has completed successfully.
*/
if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) {
nerror = hammer2_sync_replace(
thr, parent, chain,
0,
idx, focus);
dodefer = 1;
} else {
nerror = hammer2_sync_replace(
thr, parent, chain,
focus->bref.modify_tid,
idx, focus);
}
} else if (n == 0) {
/*
* 100% match, advance both
*/
advance_local = 1;
advance_xop = 1;
nerror = 0;
} else if (n > 0) {
/*
* Insert missing local data.
*
* If we will recurse, we still need to update to
* compatible content first but we do not synchronize
* modify_tid until the entire recursion has
* completed successfully.
*/
if (focus->bref.type == HAMMER2_BREF_TYPE_INODE) {
nerror = hammer2_sync_insert(
thr, &parent, &chain,
0,
idx, focus);
dodefer = 2;
} else {
nerror = hammer2_sync_insert(
thr, &parent, &chain,
focus->bref.modify_tid,
idx, focus);
}
advance_local = 1;
advance_xop = 1;
}
/*
* We cannot recurse depth-first because the XOP is still
* running in node threads for this scan. Create a placemarker
* by obtaining and record the hammer2_inode.
*
* We excluded our node from the XOP so we must temporarily
* add it to xop->head.cluster so it is properly incorporated
* into the inode.
*
* The deferral is pushed onto a LIFO list for bottom-up
* synchronization.
*/
if (error == 0 && dodefer) {
hammer2_inode_t *nip;
hammer2_deferred_ip_t *defer;
KKASSERT(focus->bref.type == HAMMER2_BREF_TYPE_INODE);
defer = kmalloc(sizeof(*defer), M_HAMMER2,
M_WAITOK | M_ZERO);
KKASSERT(xop->head.cluster.array[idx].chain == NULL);
xop->head.cluster.array[idx].flags =
HAMMER2_CITEM_INVALID;
xop->head.cluster.array[idx].chain = chain;
nip = hammer2_inode_get(pmp, ip,
&xop->head.cluster, idx);
xop->head.cluster.array[idx].chain = NULL;
hammer2_inode_ref(nip);
hammer2_inode_unlock(nip);
defer->next = list->base;
defer->ip = nip;
list->base = defer;
++list->count;
needrescan = 1;
}
/*
* If at least one deferral was added and the deferral
* list has grown too large, stop adding more. This
* will trigger an EAGAIN return.
*/
if (needrescan && list->count > 1000)
break;
/*
* Advancements for iteration.
*/
if (advance_xop) {
error = hammer2_xop_collect(&xop->head, 0);
}
if (advance_local) {
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
HAMMER2_LOOKUP_SHARED |
HAMMER2_LOOKUP_NODIRECT |
HAMMER2_LOOKUP_NODATA);
}
}
hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
/*
* If we added deferrals we want the caller to synchronize them
* and then call us again.
*
* NOTE: In this situation we do not yet want to synchronize our
* inode, setting the error code also has that effect.
*/
if (error == 0 && needrescan)
error = EAGAIN;
/*
* If no error occurred and work was performed, synchronize the
* inode meta-data itself.
*
* XXX inode lock was lost
*/
if (error == 0 && wantupdate) {
hammer2_xop_ipcluster_t *xop2;
hammer2_chain_t *focus;
xop2 = hammer2_xop_alloc(ip, HAMMER2_XOP_MODIFYING);
hammer2_xop_start_except(&xop2->head, hammer2_xop_ipcluster,
idx);
error = hammer2_xop_collect(&xop2->head, 0);
if (error == 0) {
focus = xop2->head.cluster.focus;
kprintf("syncthr: update inode %p (%s)\n",
focus,
(focus ?
(char *)focus->data->ipdata.filename : "?"));
chain = hammer2_inode_chain_and_parent(ip, idx,
&parent,
HAMMER2_RESOLVE_ALWAYS |
HAMMER2_RESOLVE_SHARED);
KKASSERT(parent != NULL);
nerror = hammer2_sync_replace(
thr, parent, chain,
sync_tid,
idx, focus);
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
/* XXX */
}
hammer2_xop_retire(&xop2->head, HAMMER2_XOPMASK_VOP);
}
return error;
}
/*
* Create a missing chain by copying the focus from another device.
*
* On entry *parentp and focus are both locked shared. The chain will be
* created and returned in *chainp also locked shared.
*/
static
int
hammer2_sync_insert(hammer2_thread_t *thr,
hammer2_chain_t **parentp, hammer2_chain_t **chainp,
hammer2_tid_t mtid, int idx, hammer2_chain_t *focus)
{
hammer2_chain_t *chain;
#if HAMMER2_THREAD_DEBUG
if (hammer2_debug & 1)
kprintf("insert rec par=%p/%d.%016jx slave %d %d.%016jx mod=%016jx\n",
*parentp,
(*parentp)->bref.type,
(*parentp)->bref.key,
idx,
focus->bref.type, focus->bref.key, mtid);
#endif
/*
* Create the missing chain. Exclusive locks are needed.
*
* Have to be careful to avoid deadlocks.
*/
if (*chainp)
hammer2_chain_unlock(*chainp);
hammer2_chain_unlock(*parentp);
hammer2_chain_lock(*parentp, HAMMER2_RESOLVE_ALWAYS);
/* reissue lookup? */
chain = NULL;
hammer2_chain_create(parentp, &chain, thr->pmp,
focus->bref.key, focus->bref.keybits,
focus->bref.type, focus->bytes,
mtid, 0);
hammer2_chain_modify(chain, mtid, 0);
/*
* Copy focus to new chain
*/
/* type already set */
chain->bref.methods = focus->bref.methods;
/* keybits already set */
chain->bref.vradix = focus->bref.vradix;
/* mirror_tid set by flush */
KKASSERT(chain->bref.modify_tid == mtid);
chain->bref.flags = focus->bref.flags;
/* key already present */
/* check code will be recalculated */
/*
* Copy data body.
*/
switch(chain->bref.type) {
case HAMMER2_BREF_TYPE_INODE:
if ((focus->data->ipdata.meta.op_flags &
HAMMER2_OPFLAG_DIRECTDATA) == 0) {
bcopy(focus->data, chain->data,
offsetof(hammer2_inode_data_t, u));
break;
}
/* fall through */
case HAMMER2_BREF_TYPE_DATA:
bcopy(focus->data, chain->data, chain->bytes);
hammer2_chain_setcheck(chain, chain->data);
break;
default:
KKASSERT(0);
break;
}
hammer2_chain_unlock(chain); /* unlock, leave ref */
if (*chainp)
hammer2_chain_drop(*chainp);
*chainp = chain; /* will be returned locked */
/*
* Avoid ordering deadlock when relocking.
*/
hammer2_chain_unlock(*parentp);
hammer2_chain_lock(*parentp, HAMMER2_RESOLVE_SHARED |
HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_lock(chain, HAMMER2_RESOLVE_SHARED |
HAMMER2_RESOLVE_ALWAYS);
return 0;
}
/*
* Repoint ip->cluster's chains to cluster's chains and fixup the default
* focus. All items, valid or invalid, are repointed. hammer2_xop_start()
* filters out invalid or non-matching elements.
*
* Caller must hold the inode and cluster exclusive locked, if not NULL,
* must also be locked.
*
* Cluster may be NULL to clean out any chains in ip->cluster.
*/
void
hammer2_inode_repoint(hammer2_inode_t *ip, hammer2_inode_t *pip,
hammer2_cluster_t *cluster)
{
hammer2_chain_t *dropch[HAMMER2_MAXCLUSTER];
hammer2_chain_t *ochain;
hammer2_chain_t *nchain;
int i;
bzero(dropch, sizeof(dropch));
/*
* Replace chains in ip->cluster with chains from cluster and
* adjust the focus if necessary.
*
* NOTE: nchain and/or ochain can be NULL due to gaps
* in the cluster arrays.
*/
hammer2_spin_ex(&ip->cluster_spin);
for (i = 0; cluster && i < cluster->nchains; ++i) {
/*
* Do not replace elements which are the same. Also handle
* element count discrepancies.
*/
nchain = cluster->array[i].chain;
if (i < ip->cluster.nchains) {
ochain = ip->cluster.array[i].chain;
if (ochain == nchain)
continue;
} else {
ochain = NULL;
}
/*
* Make adjustments
*/
ip->cluster.array[i].chain = nchain;
ip->cluster.array[i].flags &= ~HAMMER2_CITEM_INVALID;
ip->cluster.array[i].flags |= cluster->array[i].flags &
HAMMER2_CITEM_INVALID;
if (nchain)
hammer2_chain_ref(nchain);
dropch[i] = ochain;
}
/*
* Release any left-over chains in ip->cluster.
*/
while (i < ip->cluster.nchains) {
nchain = ip->cluster.array[i].chain;
if (nchain) {
ip->cluster.array[i].chain = NULL;
ip->cluster.array[i].flags |= HAMMER2_CITEM_INVALID;
}
dropch[i] = nchain;
++i;
}
/*
* Fixup fields. Note that the inode-embedded cluster is never
* directly locked.
*/
if (cluster) {
ip->cluster.nchains = cluster->nchains;
ip->cluster.focus = cluster->focus;
ip->cluster.flags = cluster->flags & ~HAMMER2_CLUSTER_LOCKED;
} else {
ip->cluster.nchains = 0;
ip->cluster.focus = NULL;
ip->cluster.flags &= ~HAMMER2_CLUSTER_ZFLAGS;
}
hammer2_spin_unex(&ip->cluster_spin);
/*
* Cleanup outside of spinlock
*/
while (--i >= 0) {
if (dropch[i])
hammer2_chain_drop(dropch[i]);
}
}
/*
* Drop an inode reference, freeing the inode when the last reference goes
* away.
*/
void
hammer2_inode_drop(hammer2_inode_t *ip)
{
hammer2_chain_drop(ip->hmp, &ip->chain);
}
/*
* Shift *chainp up to the specified directory, change the filename
* to "0xINODENUMBER", and adjust the key. The chain becomes the
* invisible hardlink target.
*
* The original *chainp has already been marked deleted.
*/
static
void
hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **chainp,
hammer2_inode_t *dip, hammer2_chain_t **dchainp,
int nlinks, int *errorp)
{
hammer2_inode_data_t *nipdata;
hammer2_chain_t *chain;
hammer2_chain_t *xchain;
hammer2_key_t key_dummy;
hammer2_key_t lhc;
hammer2_blockref_t bref;
int cache_index = -1;
chain = *chainp;
lhc = chain->data->ipdata.inum;
KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
/*
* Locate the inode or indirect block to create the new
* entry in. lhc represents the inode number so there is
* no collision iteration.
*
* There should be no key collisions with invisible inode keys.
*
* WARNING! Must use inode_lock_ex() on dip to handle a stale
* dip->chain cache.
*/
retry:
*errorp = 0;
xchain = hammer2_chain_lookup(dchainp, &key_dummy,
lhc, lhc, &cache_index, 0);
if (xchain) {
kprintf("X3 chain %p dip %p dchain %p dip->chain %p\n",
xchain, dip, *dchainp, dip->chain);
hammer2_chain_unlock(xchain);
xchain = NULL;
*errorp = ENOSPC;
#if 0
Debugger("X3");
#endif
}
/*
* Create entry in common parent directory using the seek position
* calculated above.
*
* We must refactor chain because it might have been shifted into
* an indirect chain by the create.
*/
if (*errorp == 0) {
KKASSERT(xchain == NULL);
#if 0
*errorp = hammer2_chain_create(trans, dchainp, &xchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,/* n/a */
HAMMER2_INODE_BYTES); /* n/a */
#endif
/*XXX this somehow isn't working on chain XXX*/
/*KKASSERT(xxx)*/
}
/*
* Cleanup and handle retries.
*/
if (*errorp == EAGAIN) {
kprintf("R");
hammer2_chain_wait(*dchainp);
hammer2_chain_drop(*dchainp);
goto retry;
}
/*
* Handle the error case
*/
if (*errorp) {
panic("error2");
KKASSERT(xchain == NULL);
return;
}
/*
* Use xchain as a placeholder for (lhc). Duplicate chain to the
* same target bref as xchain and then delete xchain. The duplication
* occurs after xchain in flush order even though xchain is deleted
* after the duplication. XXX
*
* WARNING! Duplications (to a different parent) can cause indirect
* blocks to be inserted, refactor xchain.
*/
bref = chain->bref;
bref.key = lhc; /* invisible dir entry key */
bref.keybits = 0;
hammer2_chain_duplicate(trans, dchainp, &chain, &bref, 0, 2);
/*
* chain is now 'live' again.. adjust the filename.
*
* Directory entries are inodes but this is a hidden hardlink
* target. The name isn't used but to ease debugging give it
* a name after its inode number.
*/
hammer2_chain_modify(trans, &chain, 0);
nipdata = &chain->data->ipdata;
ksnprintf(nipdata->filename, sizeof(nipdata->filename),
"0x%016jx", (intmax_t)nipdata->inum);
nipdata->name_len = strlen(nipdata->filename);
nipdata->name_key = lhc;
nipdata->nlinks += nlinks;
*chainp = chain;
}
/*
* Unlink the file from the specified directory inode. The directory inode
* does not need to be locked.
*
* isdir determines whether a directory/non-directory check should be made.
* No check is made if isdir is set to -1.
*
* isopen specifies whether special unlink-with-open-descriptor handling
* must be performed. If set to -1 the caller is deleting a PFS and we
* check whether the chain is mounted or not (chain->pmp != NULL). 1 is
* implied if it is mounted.
*
* If isopen is 1 and nlinks drops to 0 this function must move the chain
* to a special hidden directory until last-close occurs on the file.
*
* NOTE! The underlying file can still be active with open descriptors
* or if the chain is being manually held (e.g. for rename).
*
* The caller is responsible for fixing up ip->chain if e.g. a
* rename occurs (see chain_duplicate()).
*/
int
hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
const uint8_t *name, size_t name_len,
int isdir, int *hlinkp, struct nchandle *nch)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *parent;
hammer2_chain_t *ochain;
hammer2_chain_t *chain;
hammer2_chain_t *dparent;
hammer2_chain_t *dchain;
hammer2_key_t key_dummy;
hammer2_key_t key_next;
hammer2_key_t lhc;
int error;
int cache_index = -1;
uint8_t type;
error = 0;
ochain = NULL;
lhc = hammer2_dirhash(name, name_len);
/*
* Search for the filename in the directory
*/
if (hlinkp)
*hlinkp = 0;
parent = hammer2_inode_lock_ex(dip);
chain = hammer2_chain_lookup(&parent, &key_next,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index, 0);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
name_len == chain->data->ipdata.name_len &&
bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
break;
}
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next,
lhc + HAMMER2_DIRHASH_LOMASK,
&cache_index, 0);
}
hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
/*
* Not found or wrong type (isdir < 0 disables the type check).
* If a hardlink pointer, type checks use the hardlink target.
*/
if (chain == NULL) {
error = ENOENT;
goto done;
}
if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
if (hlinkp)
*hlinkp = 1;
type = chain->data->ipdata.target_type;
}
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
error = ENOTDIR;
goto done;
}
if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
error = EISDIR;
goto done;
}
/*
* Hardlink must be resolved. We can't hold the parent locked
* while we do this or we could deadlock.
*
* On success chain will be adjusted to point at the hardlink target
* and ochain will point to the hardlink pointer in the original
* directory. Otherwise chain remains pointing to the original.
*/
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
hammer2_chain_unlock(parent);
parent = NULL;
error = hammer2_hardlink_find(dip, &chain, &ochain);
}
/*
* If this is a directory the directory must be empty. However, if
* isdir < 0 we are doing a rename and the directory does not have
* to be empty, and if isdir > 1 we are deleting a PFS/snapshot
* and the directory does not have to be empty.
*
* NOTE: We check the full key range here which covers both visible
* and invisible entries. Theoretically there should be no
* invisible (hardlink target) entries if there are no visible
* entries.
*/
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
dparent = hammer2_chain_lookup_init(chain, 0);
dchain = hammer2_chain_lookup(&dparent, &key_dummy,
0, (hammer2_key_t)-1,
&cache_index,
HAMMER2_LOOKUP_NODATA);
if (dchain) {
hammer2_chain_unlock(dchain);
hammer2_chain_lookup_done(dparent);
error = ENOTEMPTY;
goto done;
}
hammer2_chain_lookup_done(dparent);
dparent = NULL;
/* dchain NULL */
}
/*
* Ok, we can now unlink the chain. We always decrement nlinks even
* if the entry can be deleted in case someone has the file open and
* does an fstat().
*
* The chain itself will no longer be in the on-media topology but
* can still be flushed to the media (e.g. if an open descriptor
* remains). When the last vnode/ip ref goes away the chain will
* be marked unmodified, avoiding any further (now unnecesary) I/O.
*
* A non-NULL ochain indicates a hardlink.
*/
if (ochain) {
/*
* Delete the original hardlink pointer unconditionally.
* (any open descriptors will migrate to the hardlink
* target and have no affect on this operation).
*
* NOTE: parent from above is NULL when ochain != NULL
* so we can reuse it.
*/
hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_delete(trans, ochain, 0);
hammer2_chain_unlock(ochain);
}
/*
* Decrement nlinks on the hardlink target (or original file if
* there it was not hardlinked). Delete the target when nlinks
* reaches 0 with special handling if (isopen) is set.
*
* NOTE! In DragonFly the vnops function calls cache_unlink() after
* calling us here to clean out the namecache association,
* (which does not represent a ref for the open-test), and to
* force finalization of the vnode if/when the last ref gets
* dropped.
*
* NOTE! Files are unlinked by rename and then relinked. nch will be
* passed as NULL in this situation. hammer2_inode_connect()
* will bump nlinks.
*/
KKASSERT(chain != NULL);
hammer2_chain_modify(trans, &chain, 0);
ipdata = &chain->data->ipdata;
--ipdata->nlinks;
if ((int64_t)ipdata->nlinks < 0) /* XXX debugging */
ipdata->nlinks = 0;
if (ipdata->nlinks == 0) {
if ((chain->flags & HAMMER2_CHAIN_PFSROOT) && chain->pmp) {
error = EINVAL;
kprintf("hammer2: PFS \"%s\" cannot be deleted "
"while still mounted\n",
ipdata->filename);
goto done;
}
if (nch && cache_isopen(nch)) {
kprintf("WARNING: unlinking open file\n");
atomic_set_int(&chain->flags, HAMMER2_CHAIN_UNLINKED);
hammer2_inode_move_to_hidden(trans, &chain,
ipdata->inum);
} else {
hammer2_chain_delete(trans, chain, 0);
}
}
error = 0;
done:
if (chain)
hammer2_chain_unlock(chain);
if (parent)
hammer2_chain_lookup_done(parent);
if (ochain)
hammer2_chain_drop(ochain);
return error;
}
/*
* Unlink the file from the specified directory inode. The directory inode
* does not need to be locked.
*
* isdir determines whether a directory/non-directory check should be made.
* No check is made if isdir is set to -1.
*
* NOTE! This function does not prevent the underlying file from still
* being used if it has other refs (such as from an inode, or if it's
* chain is manually held). However, the caller is responsible for
* fixing up ip->chain if e.g. a rename occurs (see chain_duplicate()).
*/
int
hammer2_unlink_file(hammer2_trans_t *trans, hammer2_inode_t *dip,
const uint8_t *name, size_t name_len,
int isdir, int *hlinkp)
{
hammer2_inode_data_t *ipdata;
hammer2_chain_t *parent;
hammer2_chain_t *ochain;
hammer2_chain_t *chain;
hammer2_chain_t *dparent;
hammer2_chain_t *dchain;
hammer2_key_t lhc;
int error;
uint8_t type;
error = 0;
ochain = NULL;
lhc = hammer2_dirhash(name, name_len);
/*
* Search for the filename in the directory
*/
if (hlinkp)
*hlinkp = 0;
parent = hammer2_inode_lock_ex(dip);
chain = hammer2_chain_lookup(&parent,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
0);
while (chain) {
if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
name_len == chain->data->ipdata.name_len &&
bcmp(name, chain->data->ipdata.filename, name_len) == 0) {
break;
}
chain = hammer2_chain_next(&parent, chain,
lhc, lhc + HAMMER2_DIRHASH_LOMASK,
0);
}
hammer2_inode_unlock_ex(dip, NULL); /* retain parent */
/*
* Not found or wrong type (isdir < 0 disables the type check).
* If a hardlink pointer, type checks use the hardlink target.
*/
if (chain == NULL) {
error = ENOENT;
goto done;
}
if ((type = chain->data->ipdata.type) == HAMMER2_OBJTYPE_HARDLINK) {
if (hlinkp)
*hlinkp = 1;
type = chain->data->ipdata.target_type;
}
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 0) {
error = ENOTDIR;
goto done;
}
if (type != HAMMER2_OBJTYPE_DIRECTORY && isdir >= 1) {
error = EISDIR;
goto done;
}
/*
* Hardlink must be resolved. We can't hold parent locked while we
* do this or we could deadlock.
*
* On success chain will be adjusted to point at the hardlink target
* and ochain will point to the hardlink pointer in the original
* directory. Otherwise chain remains pointing to the original.
*/
if (chain->data->ipdata.type == HAMMER2_OBJTYPE_HARDLINK) {
hammer2_chain_unlock(parent);
parent = NULL;
error = hammer2_hardlink_find(dip, &chain, &ochain);
}
/*
* If this is a directory the directory must be empty. However, if
* isdir < 0 we are doing a rename and the directory does not have
* to be empty, and if isdir > 1 we are deleting a PFS/snapshot
* and the directory does not have to be empty.
*
* NOTE: We check the full key range here which covers both visible
* and invisible entries. Theoretically there should be no
* invisible (hardlink target) entries if there are no visible
* entries.
*/
if (type == HAMMER2_OBJTYPE_DIRECTORY && isdir == 1) {
dparent = hammer2_chain_lookup_init(chain, 0);
dchain = hammer2_chain_lookup(&dparent,
0, (hammer2_key_t)-1,
HAMMER2_LOOKUP_NODATA);
if (dchain) {
hammer2_chain_unlock(dchain);
hammer2_chain_lookup_done(dparent);
error = ENOTEMPTY;
goto done;
}
hammer2_chain_lookup_done(dparent);
dparent = NULL;
/* dchain NULL */
}
/*
* Ok, we can now unlink the chain. We always decrement nlinks even
* if the entry can be deleted in case someone has the file open and
* does an fstat().
*
* The chain itself will no longer be in the on-media topology but
* can still be flushed to the media (e.g. if an open descriptor
* remains). When the last vnode/ip ref goes away the chain will
* be marked unmodified, avoiding any further (now unnecesary) I/O.
*
* A non-NULL ochain indicates a hardlink.
*/
if (ochain) {
/*
* Delete the original hardlink pointer.
*
* NOTE: parent from above is NULL when ochain != NULL
* so we can reuse it.
*/
hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
hammer2_chain_delete(trans, ochain);
hammer2_chain_unlock(ochain);
/*
* Then decrement nlinks on hardlink target, deleting
* the target when nlinks drops to 0.
*/
hammer2_chain_modify(trans, &chain, 0);
--chain->data->ipdata.nlinks;
if (chain->data->ipdata.nlinks == 0)
hammer2_chain_delete(trans, chain);
} else {
/*
* Otherwise this was not a hardlink and we can just
* remove the entry and decrement nlinks.
*
* NOTE: *_get() integrates chain's lock into the inode lock.
*/
hammer2_chain_modify(trans, &chain, 0);
ipdata = &chain->data->ipdata;
--ipdata->nlinks;
hammer2_chain_delete(trans, chain);
}
error = 0;
done:
if (chain)
hammer2_chain_unlock(chain);
if (parent)
hammer2_chain_lookup_done(parent);
if (ochain)
hammer2_chain_drop(ochain);
return error;
}
/*
* Create a new inode in the specified directory using the vattr to
* figure out the type of inode.
*
* If no error occurs the new inode with its chain locked is returned in
* *nipp, otherwise an error is returned and *nipp is set to NULL.
*
* If vap and/or cred are NULL the related fields are not set and the
* inode type defaults to a directory. This is used when creating PFSs
* under the super-root, so the inode number is set to 1 in this case.
*
* dip is not locked on entry.
*/
hammer2_inode_t *
hammer2_inode_create(hammer2_trans_t *trans, hammer2_inode_t *dip,
struct vattr *vap, struct ucred *cred,
const uint8_t *name, size_t name_len,
hammer2_chain_t **chainp, int *errorp)
{
hammer2_inode_data_t *dipdata;
hammer2_inode_data_t *nipdata;
hammer2_chain_t *chain;
hammer2_chain_t *parent;
hammer2_inode_t *nip;
hammer2_key_t key_dummy;
hammer2_key_t lhc;
int error;
uid_t xuid;
uuid_t dip_uid;
uuid_t dip_gid;
uint32_t dip_mode;
uint8_t dip_algo;
int cache_index = -1;
lhc = hammer2_dirhash(name, name_len);
*errorp = 0;
/*
* Locate the inode or indirect block to create the new
* entry in. At the same time check for key collisions
* and iterate until we don't get one.
*
* NOTE: hidden inodes do not have iterators.
*/
retry:
parent = hammer2_inode_lock_ex(dip);
dipdata = &dip->chain->data->ipdata;
dip_uid = dipdata->uid;
dip_gid = dipdata->gid;
dip_mode = dipdata->mode;
dip_algo = dipdata->comp_algo;
error = 0;
while (error == 0) {
chain = hammer2_chain_lookup(&parent, &key_dummy,
lhc, lhc, &cache_index, 0);
if (chain == NULL)
break;
if ((lhc & HAMMER2_DIRHASH_VISIBLE) == 0)
error = ENOSPC;
if ((lhc & HAMMER2_DIRHASH_LOMASK) == HAMMER2_DIRHASH_LOMASK)
error = ENOSPC;
hammer2_chain_unlock(chain);
chain = NULL;
++lhc;
}
if (error == 0) {
error = hammer2_chain_create(trans, &parent, &chain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,
HAMMER2_INODE_BYTES);
}
/*
* Cleanup and handle retries.
*/
if (error == EAGAIN) {
hammer2_chain_ref(parent);
hammer2_inode_unlock_ex(dip, parent);
hammer2_chain_wait(parent);
hammer2_chain_drop(parent);
goto retry;
}
hammer2_inode_unlock_ex(dip, parent);
if (error) {
KKASSERT(chain == NULL);
*errorp = error;
return (NULL);
}
/*
* Set up the new inode.
*
* NOTE: *_get() integrates chain's lock into the inode lock.
*
* NOTE: Only one new inode can currently be created per
* transaction. If the need arises we can adjust
* hammer2_trans_init() to allow more.
*
* NOTE: nipdata will have chain's blockset data.
*/
chain->data->ipdata.inum = trans->inode_tid;
nip = hammer2_inode_get(dip->pmp, dip, chain);
nipdata = &chain->data->ipdata;
if (vap) {
KKASSERT(trans->inodes_created == 0);
nipdata->type = hammer2_get_obj_type(vap->va_type);
nipdata->inum = trans->inode_tid;
++trans->inodes_created;
switch (nipdata->type) {
case HAMMER2_OBJTYPE_CDEV:
case HAMMER2_OBJTYPE_BDEV:
nipdata->rmajor = vap->va_rmajor;
nipdata->rminor = vap->va_rminor;
break;
default:
break;
}
} else {
nipdata->type = HAMMER2_OBJTYPE_DIRECTORY;
nipdata->inum = 1;
}
/* Inherit parent's inode compression mode. */
nip->comp_heuristic = 0;
nipdata->comp_algo = dip_algo;
nipdata->version = HAMMER2_INODE_VERSION_ONE;
hammer2_update_time(&nipdata->ctime);
nipdata->mtime = nipdata->ctime;
if (vap)
nipdata->mode = vap->va_mode;
nipdata->nlinks = 1;
if (vap) {
if (dip && dip->pmp) {
xuid = hammer2_to_unix_xid(&dip_uid);
xuid = vop_helper_create_uid(dip->pmp->mp,
dip_mode,
xuid,
cred,
&vap->va_mode);
} else {
/* super-root has no dip and/or pmp */
xuid = 0;
}
if (vap->va_vaflags & VA_UID_UUID_VALID)
nipdata->uid = vap->va_uid_uuid;
else if (vap->va_uid != (uid_t)VNOVAL)
hammer2_guid_to_uuid(&nipdata->uid, vap->va_uid);
else
hammer2_guid_to_uuid(&nipdata->uid, xuid);
if (vap->va_vaflags & VA_GID_UUID_VALID)
nipdata->gid = vap->va_gid_uuid;
else if (vap->va_gid != (gid_t)VNOVAL)
hammer2_guid_to_uuid(&nipdata->gid, vap->va_gid);
else if (dip)
nipdata->gid = dip_gid;
}
/*
* Regular files and softlinks allow a small amount of data to be
* directly embedded in the inode. This flag will be cleared if
* the size is extended past the embedded limit.
*/
if (nipdata->type == HAMMER2_OBJTYPE_REGFILE ||
nipdata->type == HAMMER2_OBJTYPE_SOFTLINK) {
nipdata->op_flags |= HAMMER2_OPFLAG_DIRECTDATA;
}
KKASSERT(name_len < HAMMER2_INODE_MAXNAME);
bcopy(name, nipdata->filename, name_len);
nipdata->name_key = lhc;
nipdata->name_len = name_len;
*chainp = chain;
return (nip);
}
/*
* ochain represents the target file inode. We need to move it to the
* specified common parent directory (dip) and rename it to a special
* invisible "0xINODENUMBER" filename.
*
* We use chain_duplicate and duplicate ochain at the new location,
* renaming it appropriately. We create a temporary chain and
* then delete it to placemark where the duplicate will go. Both of
* these use the inode number for (lhc) (the key), generating the
* invisible filename.
*/
static
hammer2_chain_t *
hammer2_hardlink_shiftup(hammer2_trans_t *trans, hammer2_chain_t **ochainp,
hammer2_inode_t *dip, int *errorp)
{
hammer2_inode_data_t *nipdata;
hammer2_chain_t *parent;
hammer2_chain_t *ochain;
hammer2_chain_t *nchain;
hammer2_chain_t *tmp;
hammer2_key_t lhc;
hammer2_blockref_t bref;
ochain = *ochainp;
*errorp = 0;
lhc = ochain->data->ipdata.inum;
KKASSERT((lhc & HAMMER2_DIRHASH_VISIBLE) == 0);
/*
* Locate the inode or indirect block to create the new
* entry in. lhc represents the inode number so there is
* no collision iteration.
*
* There should be no key collisions with invisible inode keys.
*/
retry:
parent = hammer2_chain_lookup_init(dip->chain, 0);
nchain = hammer2_chain_lookup(&parent, lhc, lhc, 0);
if (nchain) {
kprintf("X3 chain %p parent %p dip %p dip->chain %p\n",
nchain, parent, dip, dip->chain);
hammer2_chain_unlock(nchain);
nchain = NULL;
*errorp = ENOSPC;
#if 1
Debugger("X3");
#endif
}
/*
* Create entry in common parent directory using the seek position
* calculated above.
*/
if (*errorp == 0) {
KKASSERT(nchain == NULL);
*errorp = hammer2_chain_create(trans, &parent, &nchain,
lhc, 0,
HAMMER2_BREF_TYPE_INODE,/* n/a */
HAMMER2_INODE_BYTES); /* n/a */
hammer2_chain_refactor(&ochain);
*ochainp = ochain;
}
/*
* Cleanup and handle retries.
*/
if (*errorp == EAGAIN) {
hammer2_chain_ref(parent);
hammer2_chain_lookup_done(parent);
hammer2_chain_wait(parent);
hammer2_chain_drop(parent);
goto retry;
}
/*
* Handle the error case
*/
if (*errorp) {
KKASSERT(nchain == NULL);
hammer2_chain_lookup_done(parent);
return (NULL);
}
/*
* Use chain as a placeholder for (lhc), delete it and replace
* it with our duplication.
*
* Gain a second lock on ochain for the duplication function to
* unlock, maintain the caller's original lock across the call.
*
* This is a bit messy.
*/
hammer2_chain_delete(trans, nchain);
hammer2_chain_lock(ochain, HAMMER2_RESOLVE_ALWAYS);
tmp = ochain;
bref = tmp->bref;
bref.key = lhc; /* invisible dir entry key */
bref.keybits = 0;
hammer2_chain_duplicate(trans, parent, nchain->index, &tmp, &bref);
hammer2_chain_lookup_done(parent);
hammer2_chain_unlock(nchain); /* no longer needed */
/*
* Now set chain to our duplicate and modify it appropriately.
*
* Directory entries are inodes but this is a hidden hardlink
* target. The name isn't used but to ease debugging give it
* a name after its inode number.
*/
nchain = tmp;
tmp = NULL; /* safety */
hammer2_chain_modify(trans, &nchain, HAMMER2_MODIFY_ASSERTNOCOPY);
nipdata = &nchain->data->ipdata;
ksnprintf(nipdata->filename, sizeof(nipdata->filename),
"0x%016jx", (intmax_t)nipdata->inum);
nipdata->name_len = strlen(nipdata->filename);
nipdata->name_key = lhc;
return (nchain);
}
/*
* Retire a XOP. Used by both the VOP frontend and by the XOP backend.
*/
void
hammer2_xop_retire(hammer2_xop_head_t *xop, uint32_t mask)
{
hammer2_xop_group_t *xgrp;
hammer2_chain_t *chain;
int i;
xgrp = xop->xgrp;
/*
* Remove the frontend or remove a backend feeder. When removing
* the frontend we must wakeup any backend feeders who are waiting
* for FIFO space.
*
* XXX optimize wakeup.
*/
KKASSERT(xop->run_mask & mask);
if (atomic_fetchadd_int(&xop->run_mask, -mask) != mask) {
if (mask == HAMMER2_XOPMASK_VOP)
wakeup(xop);
return;
}
/*
* Cleanup the collection cluster.
*/
for (i = 0; i < xop->cluster.nchains; ++i) {
xop->cluster.array[i].flags = 0;
chain = xop->cluster.array[i].chain;
if (chain) {
xop->cluster.array[i].chain = NULL;
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
}
/*
* Cleanup the fifos, use check_counter to optimize the loop.
*/
mask = xop->chk_mask;
for (i = 0; mask && i < HAMMER2_MAXCLUSTER; ++i) {
hammer2_xop_fifo_t *fifo = &xop->collect[i];
while (fifo->ri != fifo->wi) {
chain = fifo->array[fifo->ri & HAMMER2_XOPFIFO_MASK];
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
++fifo->ri;
if (fifo->wi - fifo->ri < HAMMER2_XOPFIFO / 2)
wakeup(xop); /* XXX optimize */
}
mask &= ~(1U << i);
}
/*
* The inode is only held at this point, simply drop it.
*/
if (xop->ip) {
hammer2_inode_drop(xop->ip);
xop->ip = NULL;
}
if (xop->ip2) {
hammer2_inode_drop(xop->ip2);
xop->ip2 = NULL;
}
if (xop->ip3) {
hammer2_inode_drop(xop->ip3);
xop->ip3 = NULL;
}
if (xop->name) {
kfree(xop->name, M_HAMMER2);
xop->name = NULL;
xop->name_len = 0;
}
if (xop->name2) {
kfree(xop->name2, M_HAMMER2);
xop->name2 = NULL;
xop->name2_len = 0;
}
objcache_put(cache_xops, xop);
}
/*
* Update LNK_SPAN state
*/
static void
hammer2_update_spans(hammer2_dev_t *hmp, kdmsg_state_t *state)
{
const hammer2_inode_data_t *ripdata;
hammer2_chain_t *parent;
hammer2_chain_t *chain;
hammer2_pfs_t *spmp;
hammer2_key_t key_next;
kdmsg_msg_t *rmsg;
size_t name_len;
int cache_index = -1;
/*
* Lookup mount point under the media-localized super-root.
*
* cluster->pmp will incorrectly point to spmp and must be fixed
* up later on.
*/
spmp = hmp->spmp;
hammer2_inode_lock(spmp->iroot, 0);
parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
chain = NULL;
if (parent == NULL)
goto done;
chain = hammer2_chain_lookup(&parent, &key_next,
HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
&cache_index,
0);
while (chain) {
if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
continue;
ripdata = &chain->data->ipdata;
kprintf("UPDATE SPANS: %s\n", ripdata->filename);
rmsg = kdmsg_msg_alloc(&hmp->iocom.state0,
DMSG_LNK_SPAN | DMSGF_CREATE,
hammer2_lnk_span_reply, NULL);
rmsg->any.lnk_span.peer_id = ripdata->meta.pfs_clid;
rmsg->any.lnk_span.pfs_id = ripdata->meta.pfs_fsid;
rmsg->any.lnk_span.pfs_type = ripdata->meta.pfs_type;
rmsg->any.lnk_span.peer_type = DMSG_PEER_HAMMER2;
rmsg->any.lnk_span.proto_version = DMSG_SPAN_PROTO_1;
name_len = ripdata->meta.name_len;
if (name_len >= sizeof(rmsg->any.lnk_span.peer_label))
name_len = sizeof(rmsg->any.lnk_span.peer_label) - 1;
bcopy(ripdata->filename,
rmsg->any.lnk_span.peer_label,
name_len);
kdmsg_msg_write(rmsg);
chain = hammer2_chain_next(&parent, chain, &key_next,
key_next, HAMMER2_KEY_MAX,
&cache_index,
0);
}
hammer2_inode_unlock(spmp->iroot);
done:
if (chain) {
hammer2_chain_unlock(chain);
hammer2_chain_drop(chain);
}
if (parent) {
hammer2_chain_unlock(parent);
hammer2_chain_drop(parent);
}
}
